Packing edge-disjoint triangles in regular and almost regular tournaments

Packing edge-disjoint triangles in regular and almost regular tournaments Islam Akaria Raphael Yuster Abstract For a tournament T, let ν 3 T denote the maximum number of pairwise edge-disjoint triangles directed cycles of length 3 in T. Let ν 3 n denote the minimum of ν 3 T ranging over all regular tournaments with n vertices n odd. We conjecture that ν 3 n = 1 + o1n / and prove that n 11.43 1 o1 ν 3n n 1 + o1 improving upon the best known upper bound of n 1 n 8 and lower bound of 11.5 1 o1. The result is generalized to tournaments where the indegree and outdegree at each vertex may differ by at most βn. Keywords: tournament, packing, fractional AMS Subject classification: 05C0, 05C70 1 Introduction A tournament is a digraph such that for every two distinct vertices u, v there is exactly one edge with ends {u, v} so, either the edge uv or vu is present, and in this paper, all tournaments are finite. An edge uv is said to leave u and enter v. The number of edges leaving a vertex is its outdegree and the number of edges entering v is its indegree. A regular tournament is a tournament with the property that the indegree and outdegree of each vertex are equal. The semidegree of a vertex is the minimum of its indegree and outdegree. Tournaments are a major object of study in combinatorics and social choice theory. However, while complete graphs are unique for each order, there are exponentially many tournaments with the same order. As perhaps the most obvious property of a complete graph is its regularity, it seems interesting to study the properties of regular tournaments, and more generally, tournament with high minimum semidegree. Department of Mathematics, University of Haifa, Haifa 3105, Israel. E mail: islam.akaria@gmail.com Department of Mathematics, University of Haifa, Haifa 3105, Israel. E mail: raphy@math.haifa.ac.il Indeed, regular 1

tournaments have been studied by several researchers, see. e.g. [7,, 10, 13, 15]. As any connected undirected graph has an Eulerian orientation if and only if every vertex is of even degree, we have that there exist regular tournaments for every odd order. Eulerian tournaments are, therefore, the same as regular tournaments. In fact, there are exponentially many non-isomorphic regular tournaments with n vertices [10]. All regular tournaments have the same number of triangles and the same number of transitive triples where a triangle is a set of three edges {xy, yz, zx} while a transitive triple is a set of three edges {xy, yz, xz}. This follows from the obvious fact that the number of transitive triples and hence triangles in any tournament is determined by the score of the tournament, which is the sorted outdegree sequence. For regular tournaments this amounts to nn 1n 3/8 transitive triples and therefore to n 3 nn 1n 3/8 = nn 1/4 triangles. Asymptotically, this means that a fraction of 1/4 of the triples are triangles while 3/4 of the triples are transitive. Throughout this paper a triangle is denoted by C 3. An edge triangle packing of an undirected graph is a set of pairwise edge-disjoint subgraphs that are isomorphic to a triangle. The study of triangle packings in graphs was initiated in the classical result of Kirkman [8] who proved that the complete graph with n vertices has a triangle packing of size nn 1/ whenever n 1, 3 mod. In other words, when n 1, 3 mod there always exists a Steiner triple system STS, which is a set of triples of [n] with the property that any pair of [n] appears in exactly one of these triples. This clearly implies that for other moduli of n there are packings with 1 o n 1n / triangles 1, and this is asymptotically tight as such packings cover 1 o n 1 n edges. In the directed case, a triangle packing of a tournament requires each subgraph to be isomorphic to C 3. Triangle packings and packings by transitive triples of digraphs have been studied by several researchers see, e.g., [4,, 1]. Nontrivial lower bounds on the number of edge-disjoint triangles give upper bounds regarding the so-called Erdős-Hajnal coefficients of tournaments see []. The number of edge-disjoint triangles is a natural measure of how non-transitive a given tournament is. For a tournament T, we denote by ν 3 T the size of a largest triangle packing. Observe that ν 3 T c 3 T /n for every tournament with n 1, 3 mod where c 3 T is the total number of triangles. This can be seen by taking a random STS of n and observing that the expected number of directed triangles in the STS is nn 1/c 3 T / n 3. In particular, this means that ν 3 T 1 o n 1n /4 for any regular tournament T with n vertices. On the other hand, we always have the trivial upper bound ν 3 T 1 o n 1n /. Let, therefore ν 3 n denote the minimum of ν 3 T ranging over all regular tournaments with n vertices assuming, of course, that n is odd. Hence, trivially n 4 1 o n1 ν 3 n n 1 o n1. 1 Here o n1 means a function that goes to zero as n goes to infinity.

While exact small values of ν 3 n are known by brute force computation, determining the asymptotic value of ν 3 n seems to be a difficult problem. The best known bounds were given in [15]: n 11.5 1 o n1 ν 3 n n 1. 8 The present paper improves both the lower bound and the upper bound. While the upper bound is improved significantly, the improvement in the lower bound is milder. Theorem 1.1 1 3 7 3 ln10 n 1 o n 1 ν 3 n n 1 + o n1. Notice that 1 3 7 3 ln 10 > 1/11.43. As explained in Section, it is natural to suspect that the construction yielding the upper bound is, in a sense, a worst case construction. Thus, we make the following conjecture. Conjecture 1.1 ν 3 n = n 1 + o n1. While the proof of the upper bound in Theorem 1.1 is different from the one in [15], the proof of the lower bound is similar in many aspects. The additional important ingredient that enables us to obtain the improved lower bound is a strengthening of Lemma 3.3 there, replaced by the significantly more involved Lemma 3.4 here, which bounds the number of triangles containing dense edges edges that appear in many triangles. We are able to extend our results to not necessarily regular tournaments. We say that a tournament is β-almost-regular or, for brevity, and slightly abusing terminology, β-regular if the indegree and outdegree at each vertex differ by at most βn, Notice that this is equivalent to saying that the minimum semidegree is at least 1 βn 1/. Thus, β = 0 coincides with regular tournaments and β = 1 coincides with the family of all tournaments. Notice that this is a very general notion as for any β > 0, a sufficiently large random tournament is almost surely β-regular. Here we no longer need to require that n has a certain parity. Generalizing the above notation, we denote by ν 3 β, n the minimum of ν 3 T ranging over all β-regular tournaments with n vertices. The following extends Theorem 1.1. Theorem 1. { 1 β ν 3 β, n min 11 + β ν 3 β, n ln 11 + 1β + 3β ν 3 β, n ln 1 + β 5 + β 3β + β 3, } 1 β n 1 + o n 1. 8 n 1 o n 1 if β 1, n 1 o n 1 if β > 1. 3

Plugging in β = 0, we note that the lower bound in this case given in Theorem 1. is slightly worse than the lower bound given in Theorem 1.1. The bounds differ by less than 0.001. The reason is that the proof of Theorem 1.1 is already quite computationally involved without adding an extra parameter namely β to its computations, making them somewhat intractable, and the gain would be marginal at the expense of the clarity of the proof we provide for Theorem 1.. On the other hand, as can be seen from the proof of Theorem 1.1 which is robust to degree deviations of on without affecting its statement, we note that Theorem 1.1 also holds for almost regular tournaments, namely tournaments whose minimum semidegree is n/ on. As in the case of the proof of Theorem 1.1, the lower bound in Theorem 1. is similar in many aspects to the lower bound proof given in [15], but with one additional important ingredient required for the generalization. This is Lemma 4.1 which gives a tight lower bound for the number of triangles in β-regular tournaments and which may be of independent interest. The rest of this paper is organized as follows. In Section, we prove the upper bound in Theorem 1.1. To this end, we need to define the fractional relaxation of the problem and consider its dual covering problem. We also prove that the upper bound we obtain cannot be improved using our construction. We explain why it is natural to suspect that this construction is the worst, and hence the justification for conjecture 1.1. We also show how to generalize the construction to β-regular tournaments and obtain the upper bound in Theorem 1.. In Section 3 we prove the lower bound in Theorem 1.1. As in [15] our main tool is a result of Haxell and Rödl [5] tailored to the directed setting in [11] connecting the fractional value of a maximum packing with its integral one. Section 4 addresses the changes needed in the statements given in Section 3 in order to apply them to the more general setting of βn-regular tournaments, resulting in the proof of the lower bound in Theorem 1.. Upper bounds.1 Fractional relaxation of packing and covering We start this section by defining the fractional relaxation of the triangle packing problem together with its dual fractional covering problem, and define the parameters ν3 n and τ 3 n that are the fractional analogue of ν 3 n and its dual, respectively. Let R + denote the set of nonnegative reals. A fractional triangle packing of a digraph G is a function ψ from the set F 3 of copies of C 3 in G to R +, satisfying e X F 3 ψx 1 for each edge e EG. Letting ψ = X F 3 ψx, the fractional triangle packing number, denoted ν3 G, is defined to be the maximum of ψ taken over all fractional triangle packings ψ. Since a triangle packing is also a fractional triangle packing by letting ψ = 1 for elements of F 3 in the packing and ψ = 0 for the other elements, we always have ν 3 G ν 3G. However, the two parameters may differ. In particular, they may differ for regular tournaments. Consider, for example, the 5- vertex regular tournament consisting of two edge-disjoint directed cycles of length 5 each. Clearly, 4

ν 3 T =. On the other hand, we may assign each of the five triangles of this tournament the value 1/ thereby obtaining a fractional triangle packing of value.5. A fractional triangle cover of a digraph G is a function φ from the set of edges EG of G to R +, satisfying e X F 3 φe 1 for each triangle X F 3. Letting φ = e EG φe, the fractional triangle cover number, denoted τ3 G, is defined to be the minimum of φ taken over all fractional triangle covers φ. It is worth mentioning that τ3 G is trivially a lower bound for its integral counterpart, τ 3 G, which is the minimum number of edges covering all triangles. Hence, it is also a trivial lower bound for the minimum feedback edge set of a tournament, which is the smallest set of edges whose removal makes the tournament acyclic. By linear programming duality, τ3 G = ν 3 G. For example, in the 5-vertex regular tournament of the previous paragraph, we may assign the value 1/ to each edge of a 5-cycle and obtain a valid fractional triangle cover of value.5.. Upper bound for regular tournaments In order to obtain a good upper bound, we must first construct a regular tournament which is as transitive as possible so that it will not be able to accommodate many pairwise edge-disjoint triangles. Naturally, any regular tournament on n vertices cannot have a transitive subset on more than n + 1/ vertices, since in such a subset the outdegree of the source would already be more than n 1/. The following regular tournament, denoted R n, does have a transitive subset on n + 1/ vertices, in fact it has many such subsets. It even has many pairs of edge-disjoint such subsets each pair sharing exactly one vertex. It is reasonable to suspect that a maximum triangle packing of R n yields the value of ν 3 n. For n odd, we define R n as follows. Its vertices are {0,..., n 1} one can view them as elements of the cyclic group Z n. Vertex i has an outgoing edge towards vertex j if and only if 1 j i mod n n 1/. Thus, if we think of the vertices as lying on a directed cycle of length n, each vertex sends outgoing edges to the n 1/ vertices following it on the cycle. Observe that R n is a regular tournament and that for any vertex i, the set of vertices {i, i + 1,..., i + n 1/} indices modulo n forms a transitive subset. We will prove that ν 3 R n n+on, which implies that ν 3 n n+on. Since, by the previous subsection, τ 3 R n = ν 3 R n ν 3 R n, it suffices to prove the following. Lemma.1 τ 3 R n n + on Proof. We consider first case where n 1 mod. We will construct a particular covering which attains the bound stated in the lemma. Define the length of an edge of R n from i to j by lengthi, j = j i mod n. We give all the edges of length 1,..., n the weight 0 i.e. φe = 0 for lengthe { 1,..., } n. To each edge e of length l > n we give the weight φe = n+1 l n. 5.

Proposition. The assignment φ is a fractional triangle cover. Proof. Let h, i, j be a triangle, without loss of generality h = 0 so the triangle is 0, i, j. First case: i { 1,..., } n. Then the other edges of the triangle must have length larger than n and hence φij = n+1 j i n and φj0 = n+1 n j n. The sum of weights of the edges of the triangle 0, i, j is: = = 0 + n j i + n n j n + 1 n + 1 n n i n n 3 n + 1 n + 1 n 3 n 1 = n n 1 = 1. n + 1 n + 1 Second case: i / { 1,..., } n. Then φ0i = n+1 i n and we have three subcases for the weight of edge ij and edge j0: the first subcase is φij = n+1 j i n and φj0 = 0, the second subcase is φij = 0 and φj0 = n+1 n j n, and the last subcase is φij = n+1 j i n 0, i, j in the three subcases: First subcase: and φj0 = n+1 = = n j n. Now we calculate the weight of the triangle n i + j i n + 1 n + 1 n j n + 1 n + 1 n + 1 = 1. n + 1 n + 0 n n 1 n 1 We used the fact that in this subcase we must have lengthj, 0 n so j n n 1/. Second subcase: n i + 0 + n n j n + 1 n + 1 n = n j + i n n 1 n 1 n + 1 n + 1 n + 1 = = 1. n + 1 Recall that in this subcase lengthi, j = j i n. Third subcase: = n + 1 n + 1 n i + j i n + 1 n n 3 = n + 1 n + n 3 n 1 n + 1 = 1. n n j

This concludes the proof of Proposition.. We calculate the value of this fractional triangle cover. Observe that only lengths between n/ +1 until n/ which is the maximum possible length of an edge by the definition of R n receive nonzero weight which is the length minus n/, normalized by multiplying it with /n + 1. Thus, φ = φe = n 1 + + 3 +... + n 1 n + 1 3 e E n n 1 3 1 + n 1 3 = n + 1 n n 1n + = n + 1 < n. Hence τ3 R n < n for n 1 mod. Now, if n 1 mod, then, as n is odd, either n 3 mod or n 5 mod. Observe that since R n is a subgraph of R n+ just delete vertices 0 and n + 1/ from R n+ to obtain a subgraph isomorphic to R n we have τ 3 R n τ 3 R n+ τ 3 R n+4. Thus, for the case n 5 mod, we have that n + 1 mod hence τ3 R n τ3 R n+ n+ = n+on. For the case n 3 mod, we have that n + 4 1 mod hence τ3 R n τ3 R n+4 n+4 = n+on. This completes the proof of Lemma.1 and hence the upper bound in Theorem 1.1. One may wonder whether the fractional cover constructed in Lemma.1 is optimal for R n. Perhaps we can do better and improve the upper bound regardless of whether one believes that R n is a worst case example. In the following lemma we show that our constructed covering is asymptotically optimal for R n. Lemma.3 νr n n on Proof. We prove this for n = k k odd. In this case we will show that we can pack exactly n = k pairwise edge-disjoint triangles. We define the packing as follows. It consists of n = k sets of triangles, denoted S 0,..., S n 1. Each set will contain k pairwise edge-disjoint triangles. Overall, the construction consists of nk = n / triangles. Furthermore, for any two sets S i, S j, their triangles are pairwise edge-disjoint. We describe S j for j = 0,..., n 1. It consists of the triangles j, j + a i mod n, j + a i + b i mod n for i = 0,..., k 1 where: b i = n 1/ 3k/ + i + / for i odd. b i = n 1/ k + i/ + 1 for i even.. 1 7

a i = k + i/ for i even. a i = 3k/ + i/ for i odd. For example, if k = hence n = 81 we have that S 0 is: {0, 18, 41, 0, 14, 4, 0, 1, 43, 0, 15, 44, 0, 0, 45, 0, 1, 4, 0, 1, 47, 0, 17, 48, 0,, 4}. We need to prove that each of the listed triples in each of the S j is indeed a directed triangle of R n, and that no edge repeats twice in any of the S j. Each triple is of the form j, j + a i mod n, j + a i + b i mod n. The lengths of the edges in this triangle are a i, b i and c i = n a i b i. Observe that a i is always between 1 and n 1/ by its definition. Indeed, if i is even, then If i is odd, then n = k a i k + k 1 = 5k 1 = 5n 18 1 n 1. n + 1 = 3k + 1 a i 3k + k = 4k = n 1 n 1. In any case, the first edge of each triangle whose length is a i, is indeed an edge of R n. Observe similarly that b i is always between 1 and n 1/ by its definition. Indeed, if i is even, then If i is odd, then 5n 18 + 1 = n 1 k + 1 b i n 1 k + 1 + k 1 = n 3 n 1. n 3 + 1 = n 1 3k + 3 b i n 1 3k + k = 7n 18 1 n 1. In any case, the second edge of each triangle whose length is b i, is indeed an edge of R n. Finally, c i is always between 1 and n 1/ since by the definitions of a i and b i we have c i = n 1/ i. We have proved that each triple in each S j is a directed triangle of R n. Observe also that the interval of values of the a i is always between n/ + 1/ and 5n/18 1/. The interval of values of the b i is always between 5n/18 + 1/ and 7n/18 1/. The interval of values of the c i is always between 7n/18 + 1/ and n 1/. As these three intervals are disjoint, this proves that no edge is repeated twice in the construction. This proves the lemma when n = k and k is odd. Now, for any other odd number n, let k be the largest odd number such that k n. Recalling that R k is a subgraph of R n we have that νr n νr k k = n on. 8

.3 Upper bound for β-regular tournaments In this subsection we prove the upper bound for ν 3 β, n given in Theorem 1.. Consider the regular tournament graph R 1+βn defined in the previous subsection. We can assume 1 + βn is an odd integer as rounding issues do not affect the asymptotic claim. Delete from R 1+βn the vertices {0, 1,..., βn 1} and denote the resulting tournament by T. Notice that T has n vertices and since R n is regular, and we have removed only βn vertices from it, we have that T is a β-regular tournament. We first consider the case where β 1/5. Let φ be the fractional triangle cover defined on R 1+βn, proved in 1 to satisfy φ 1 + o n 11 + β n /. Let φ be the fractional triangle cover of T induced by φ. Namely, each edge of T retains its weight under φ. Now, φ φ is just the sum of the weights of the edges incident with the removed vertices {0, 1,..., βn 1}. By 1, the sum of the weights of the edges leaving each vertex of R 1+βn is 1 o n 11 + βn/ and, by symmetry, the sum of the weights of the edges entering each vertex of R 1+βn is also 1 o n 11 + βn/. Now, for all β 1/5 we have that βn 1 + βn/. Hence all the edges ij where i, j {0, 1,..., βn 1} have φij = 0. Thus, It follows that φ φ βn 1 o n 1 1 + β n. φ β1 + β φ 1 o n 1 1 + o n 1 1 + o n 1 1 β n. n 1 + β n 1 o n 1 β1 + β n Since ν 3 β, n ν 3 T ν 3 T = τ 3 T φ we have that ν 3 β, n 1 + o n 11 β n / for β 1/5. The following triangle cover, denoted φ is valid for all β < 1. Assign the weight 1 to all the edges of T of the form ij where i > j. All other edges receive the weight 0. Notice that each directed triangle must contain an edge having weight 1 and hence φ is a valid triangle cover in fact, an integral cover. We count the number of edges receiving weight 1. Vertex 1 + βn 1 the vertex with largest index has an outgoing edge in R 1+βn to all vertices j with j < 1 + βn/. Hence, it has at most 1 + βn/ βn 1 = n/ βn/ 1 edges leaving it in T having weight 1. Similarly, for all k = 1,..., n/ βn/, vertex 1 + βn k has at most n/ βn/ k edges leaving it in T having weight 1. Hence, φ n/ βn/ k=1 n/ βn/ k 1 + o n 1 1 β n. 8

Since ν 3 β, n ν 3 T ν 3 T = τ 3 T φ we have that ν 3 β, n 1 + o n 11 β n /8 for β 1. Observe that for all β 1/17 1/5 the bound obtained via φ is better than the bound obtained via φ hence we may summarize that { 1 β ν 3 β, n min, } 1 β n 1 + o n 1. 8 3 A lower bound for regular tournaments 3.1 Integer versus fractional packings A result of Nutov and Yuster [11] asserts that the integral and fractional parameters differ by on. The following is a very spacial case of their result. Theorem 3.1 If T is an n-vertex tournament, then ν 3 T ν 3T = on. An undirected version of Theorem 3.1 has been proved by Haxell and Rödl [5] who were the first to prove this interesting relationship between integral and fractional packings. The proof of Theorem 3.1 makes use of the directed version of Szemerédi s regularity lemma [14] that has been used implicitly in [3] and proved in [1]. Let ν3 n be the minimum of ν 3 T ranging over all n-vertex regular tournaments T. Similarly, let ν3 β, n be the minimum of ν 3 T ranging over all n-vertex β-regular tournaments T. By Theorem 3.1 and the fact that fractional packings are at least as large as integral packings we have: Corollary 3. ν 3 n ν 3n ν 3 n on. Similarly, ν 3 β, n ν 3β, n ν 3 β, n on. 3. Proof of the lower bound in Theorem 1.1 In this section we prove the following theorem that, together with Corollary 3., yields the lower bound in Theorem 1.1. Theorem 3.3 A regular tournament T with n vertices has ν3 T 1 o n1 1 3 7 3 ln 10 n. As in [15], we call an edge α-dense if it is contained in at least αn triangles. Observe that no edge is 1/-dense as any edge of a regular tournament appears in at most n 1/ triangles. We require the following lemma that bounds the number of triangles that contain α-dense edges where α is relatively large. It is an improvement over Lemma 3.3 in [15]. Lemma 3.4 For all α 1/4, the number of triangles that contain α-dense edges is at most 1 α 5 3 α 1 3 n3. Proof. As shown in [15], the total number of α-dense edges entering each vertex is at most n1 α. We repeat the details of this observation for completeness. For a vertex v, we compute the number of α-dense edges entering it. Let B v N v be the set of vertices x such that xv is 10

α-dense. Consider a vertex x of maximum indegree in the sub-tournament T [B v ] induced by B v. Since in any tournament with B v vertices the maximum indegree is at least B v 1/ we have that x has at least B v 1/ edges entering it in T [B v ]. On the other hand, as xv is α-dense, we also have that x has at least αn vertices of N + v entering it. Since N + v B v = we have that the indegree of x in T is at least B v 1/ + αn. But the indegree of x in T is n 1/ and thus B v 1/ + αn n 1/. It follows that B v n1 α. Similarly, if C v N + v is the set of vertices x such that vx is α-dense, we have that C v n1 α. But we are not interested in counting the number of α-dense edges incident with a vertex, rather we wish to count the number of triangles containing α-dense edges. To this end, we need to define certain parameters. 1. Let rv denote the number of triangles containing v in which the edge entering v is α-dense and the edge leaving v is not α-dense.. Let sv denote the number of triangles containing v in which the edge leaving v is α-dense and the edge entering v is not α-dense. 3. Let tv denote the number of triangles containing v in which both edges incident to v are α-dense. 4. Let bv = rv + tv denote the number of triangles containing v in which the edge entering v is α-dense. 5. Let cv = sv + tv denote the number of triangles containing v in which the edge leaving v is α-dense.. Let qv = 1 rv + 1 sv + 1 3 tv. We claim that v V qv is an upper bound for the total number of triangles containing an α-dense edge. Indeed, consider some triangle x, y, z containing an α-dense edge. If it contains a single α-dense edge, say xy, then this triangle is counted 1/ for sx and 1/ for ry. If it contains three α-dense edges, then it is counted 1/3 for each of tx, ty, tz. If it contains precisely two α-dense edges, say xy and yz, then it is counted 1/ for sx, 1/ for rz and 1/3 for ty, so it contributes more than 1. In any case, each triangle containing an α-dense edge contributes at least 1 to the sum v V qv. It remains to upper bound v V qv. We will upper bound each qv separately, and multiply the bound by n. Notice that by the definitions of bv and cv, qv = 1 bv + 1 cv tv. 3 11

Let βn = B v and γn = C v and recall that β 1 α and γ 1 α. We start by giving upper bounds for bv and cv in terms of β and γ respectively. For any x B v, let fx denote the number of triangles containing the α-dense edge xv. By the definition of B v we have that fx αn. Let dx denote the indegree of x in T [B v ]. As in the argument at the beginning of the proof, we have that dx + fx n 1/. Now, by the definition of bv we have that bv = x B v fx and therefore bv = fx n 1 dx. x B v x B v On the other hand, x B v dx = B v B v 1/. Hence, Analogously, we have that bv B v n 1 B v B v 1 = β1 β n. 3 cv C v n 1 C v C v 1 = γ1 γ n. 4 We next give a lower bound for tv. Consider any edge xy that goes from C v to B v. This means that v, x, y is a triangle where both yv and vx are α-dense. Hence, this triangle contributes to tv. Thus, the number of edges going from C v to B v is equal to tv. There are at least B v αn edges going from N + v to B v. At most N + v C v B v of them go from N + v \ C v to B v. Hence, tv B v αn N + v C v B v = αβn n 1 γnβn βα 1 + γn. We can similarly estimate tv by the fact that there are at least C v αn edges going from C v to N v. At most N v B v C v of them go from C v to N v \ B v. Hence, tv C v αn N v B v C v = αγn n 1 Using the last two inequalities we obtain that tv βγ 1 βnγn γα 1 + βn. αβ + γ n. 5 By, 3, 4, 5 we get that qv 51 α3 β + γ β + γ 4 βγ n. Hence, our remaining task is to maximize the expression 5 1 α β+γ 3 β + γ 4 βγ subject to the constraints 0 β 1 α and 0 γ 1 α and recall that α 1/. Simple analysis of the 1

partial derivatives show that for all α 3/8, the maximum is obtained when β = γ = 1 α. When 1/4 α 3/8 the bound in the statement of the lemma trivially holds as 1 α 5 3 α 1 3 1/4 in this range and recall that a regular tournament has less than n 3 /4 triangles. Thus, in any case, plugging in β = γ = 1 α in and rearranging the terms we obtain that qv 1 α 5 3 α 1 3 n. Consequently, for all α 1/4, the number of triangles that contain α-dense edges is at most qv 1 α 5 3 α 1 3 n3. v V This finishes the proof of Lemma 3.4. For an edge e let fe denote the number of triangles that contain e. We define a fractional triangle packing ψ as in [15] by assigning to a triangle X the value ψx = 1 max e X fe. 7 In other words, we consider the three edges of X and take the edge e with fe maximal, setting ψx to 1/fe. Notice that ψ is a valid fractional triangle packing. Indeed, the sum of the weights of triangles containing any edge e is at most fe fe 1 = 1. Proof of Theorem 3.3: Let k be a positive integer, and let 1 > x > 3/4 be a parameter to be chosen later. Define c = x 1/k+1 and let α i = 1 ci+1 for i = 0,..., k. Observe that α k = x/ so 1/ > α i α k > 3/8. Define as in [15] E i = {e ET : fe α i n}. So, E i is the set of all α i -dense edges and notice that E 0 E 1 E k. For i = 0,..., k, let S i denote the set of all triangles that contain an edge from E i and do not contain an edge from E j where j < i. In particular, S 0 is just the set of triangles that contain an edge from E 0. Finally, let S k+1 be the triangles that are not in k i=0 S i and observe that S 0,..., S k+1 is a partition of the set of all nn 1/4 triangles of T. For i = 0,..., k, all the elements of S 0 S i contain edges that are α i -dense and therefore by Lemma 3.4 we have that for i = 0,..., k: t i = i j=0 S j 1 α i 5 3 α i 1 3 n3. 8 By the definition of t i we have that for i = 1,..., k, S i = t i t i 1 and that S 0 = t 0. Thus, we also have that S k+1 = nn 1 4 t k. 13

For i = 1,..., k + 1, all the elements of S i receive weight that is greater than 1/α i 1 n. Indeed, consider X S i. We know that it does not contain an edge from E j for j < i. So the maximum value of fe for an edge e of X is smaller than α i 1 n. By the definition of ψ we therefore have that ψx > 1/α i 1 n. For elements X S 0 we use the trivial bound ψx > /n. Summing up the weights of all the triangles of T we find that: ψ t 0 k n + 1 nn t i t i 1 α i 1 n + 1 4 Rearranging the terms we have: ψ n 1 t 0 4α k n 1 α 0 Using 8 we have that: ψ n 1 n 1 α 0 5 4α k 3 α 0 1 1 3 α 0 Thus, we must choose k and x so as to maximize 1 1 α 0 5 4α k 3 α 0 1 1 3 α 0 Recalling that a i /a i 1 = c the last expression is identical to k k k t k 1 α k n. t i 1 1 n α i α i 1 1 + 1 3 + 8α 0 0 4α k 3α 0 3 α 0 + 1 1 7 3α k 3α 0 3 k + 7 ck + 10 3 3. 10 n 1 α i 5 3 α i 1 1 3 1 α i α i 1 1 α i 5 3 α i 1 1 3 1 α i α i 1 Since kα i = 0.5c c k 1/c 1 the last expression is identical to k α i 10 k 3 c α i. 3 4c k+1 3 + 4c 5 3 c + 7 3 kc 1 5 3 c c k 1. Finally, recalling that c = x 1/k+1, the last expression is identical to 3 4x 3 + 4x1/k+1 5 3 x/k+1 + 7 3 kx1/k+1 1 5 3 x/k+1 x k/k+1 1. Taking the limit of the last expression as k we obtain 3 4x + 1 + 7 3 ln x 5 3 x. The maximum of the last expression for 1 > x > 3/4 is obtained at x = /10 in which case the expression amounts to This proves that ψ 1 3 7 3 ln10. 1 3 7 3 ln10 n 1 o n 1... 14

4 Lower bound for β-regular tournaments In order to generalize the lower bound for β-regular tournaments we need to address three issues. The first is that the number of triangles in β-regular tournaments may not be the same for all such tournaments, unlike regular tournaments which all have precisely nn 1/4 triangles, and we must therefore determine a tight lower bound in terms of β. The second issue requires an analogue of Lemma 3.4 suitable for β-regular tournaments. The third issue concerns the analysis of the fractional packing, generalizing the one given in the proof of Theorem 3.3. We start with a lower bound for the number of triangles in β-regular tournaments. Lemma 4.1 The number of C 3 in a β-regular tournament with n vertices is at least 1 3β 4 n 3 1 o n 1 for β 1/ and at least 1 β3 1 n 3 1 o n 1 for β > 1/. This is asymptotically tight for all 0 β 1. Proof. The number of transitive triples and hence the number of triangles in any tournament is determined by the outdegrees of the vertices. Let d i denote the outdegree of vertex i in a tournament with vertices 1,..., n. The number of transitive triples is clearly n di. and we wish to maximize this amount. In β-regular tournaments we have the additional restriction that n1 β/ d i n1 + β/. Now, suppose the degrees are sorted so that d i d i+1 for i = 1,..., n 1. In order for the tournament to be realized we have the further restriction that d 1 +... + d i i since already the first i vertices induce a tournament whose outdegree sum is i. Similarly, n 1 dn i+1 +... + n 1 d n i since already the last i vertices induce a tournament whose indegree sum is i. As the statement of the lemma is asymptotic, it is more convenient to formulate the analogous continuous convex optimization problem. maximize s.t. 1 fx dx 0 fx is monotone nondecreasing 1 β fx 1 + β α 0 1 α fxdx α 1 fxdx 1 α. When β 1/ the obvious solution, by convexity, is obtained by setting fx = 1 β/ for 0 x 1/ and fx = 1 + β/ for 1/ x 1. Observe that since β 1/, the last two 15

restrictions of the convex minimization problem trivially hold. In this case we obtain that and correspondingly, n 1 0 fx dx = 1 + β 8 di 1 + β n 3 1 + o n 1. 8 The number of triangles is therefore always at least 1 1 + β 1 3β n 3 1 o n 1 = n 3 1 o n 1. 8 4 When β > 1/, the last two restrictions of the convex minimization problem force fx to linearly increase in the range 1 β x β and we obtain the optimal solution In this case we obtain that 1 0 fx dx = and correspondingly, 1 β 1 β + 8 n fx = x 1 β 0 x 1 β 1 β < x < β 1+β β x 1. 1 + β 1 β + β3 1 β3 8 di 1 1 + 1 4 β 1 4 β + 1 1 β3 n 3 1 + o n 1. The number of triangles is therefore always at least = 1 1 + 1 4 β 1 4 β + 1 1 β3 1 1 1 4 β + 1 4 β 1 1 β3 n 3 1 o n 1 = 1 β3 n 3 1 o n 1. 1 The result is asymptotically tight for every β as the extremal degree sequences are realizable as β-regular tournaments. For β 1/ we can take two disjoint regular tournaments A and B on n/ vertices each. We can then take 1/4 β/n disjoint perfect matchings between A and B and direct all edges of these matchings from A to B. The remaining edges between A and B are directed from B to A. In the resulting tournament, each vertex of A has outdegree n1 β/ 1/ and each vertex of B has outdegree n1 + β/ 1/, hence a β-regular tournament realizing the extremal degree sequence. For β > 1/ we can take two disjoint regular tournaments A and B on βn vertices each, and an additional set of vertices denoted as x 1,..., x n1 β. Now, for i = 1,..., n1 β, direct edges from x i to all vertices of A and to all vertices x j with j < i. Direct edges to x i from all 1

vertices of B and from all vertices x j with j > i. Also direct all edges from B to A. The resulting tournament has n vertices, is β-regular, and its degree sequence realizes the extremal case. We next need to obtain an analogue of Lemma 3.4 that applies to β-regular tournaments. Although it is possible to generalize Lemma 3.4 directly, the already involved analysis become less tractable. We settle for a somewhat simpler version with only a small loss in the upper bound. Lemma 4. Let T be a β-regular tournament with n vertices. For all 0 < α < 1 + β/, the number of triangles of T that contain α-dense edges is at most n3 1+β α. Proof. For a vertex v, we compute the number of α-dense edges entering it. Let B v N v be the set of vertices x such that xv is α-dense. Consider a vertex x of maximum indegree in the sub-tournament T [B v ] induced by B v. Since in any tournament with B v vertices the maximum indegree is at least B v 1/ we have that x has at least B v 1/ edges entering it in T [B v ]. On the other hand, as xv is α-dense, we also have that x has at least αn vertices of N + v entering it. Since N + v B v = we have that the indegree of x in T is at least B v 1/ + αn. But the indegree of x in T is at most n1 + β 1/ and thus B v 1/ + αn n1 + β 1/. It follows that B v n1 + β α. Similarly, if C v N + v is the set of vertices y such that vy is α-dense, we have that C v n1 + β α. Now, each x B v lies in at most N + v triangles and each y C v lies in at most N v triangles. We therefore have that the number of triangles containing v and an α-dense edge incident with v either entering v or leaving v is at most n1 + β α N + v + N v < n 1 + β α. Summing this value for each v V and observing that each triangle that contains an α-dense edge is counted at least twice, we obtain that the number of triangles containing α-dense edges is at most n 3 1 + β α/. Finally, we need to generalize the analysis given in the proof of Theorem 3.3. We use the exact same fractional packing ψ defined in 7. As in the proof of Theorem 3.3 we let k be a positive integer, let x < 1 be a parameter to be chosen later, define c = x 1/k+1 and define α i = 1+βc i+1 / for i = 0,..., k. By Lemma 4., the upper bound for t i given in 8 is replaced with: t i = i 1 + β α j=0 S j n 3. 11 Similarly, using Lemma 4.1, the lower bound for S k+1 given in is replaced with: S k+1 1 3β n 3 1 o n 1 t k if β 1 4, S k+1 As in 10 we have, after rearranging the terms: ψ 1 3β n 1 o n 1 t 0 1 4α k n α 0 1 + β ψ 1 β3 n 1 o n 1 t 0 1 1α k n α 0 1 + β 17 1 β3 n 3 1 o n 1 t k if β > 1 1. k k t i 1 1 n α i α i 1 t i 1 1 n α i α i 1 if β 1, if β > 1.

Using 11 we have that: ψ 1 3β 4α k n 1 o n 1 n 1+β α 0 1 α0 1+β ψ k n 1+β α i 1 αi 1 1 β3 1α k n 1 o n 1 n 1+β α 0 k n 1+β α i α i 1 1 α0 1+β 1 αi 1 α i 1 if β 1, if β > 1. Thus, we must choose k and x so as to maximize ψ 1 3β 1 + β α 0 1 k 4α k α 0 1 + β 1 β3 ψ 1 + β α 0 1 k 1α k α 0 1 + β 1 + β α i n 1 + β α i 1 1 α i α i 1 1 1 α i α i 1 if β 1, if β > 1. Recalling that a i /a i 1 = c the last expression is identical to 11 1β 3β 4α k + α 0 + k1 c 1 + β if β 1/, 5 β + 3β β 3 1α k + α 0 + k1 c 1 + β if β > 1/. Recalling that c = x 1/k+1, α 0 = 1 + βc/, α k = 1 + βc k+1 / we obtain that 11 1β 3β 1x1 + β 5 β + 3β β 3 x1 + β + x 1/k+1 + k1 x 1/k+1 if β 1, + x 1/k+1 + k1 x 1/k+1 if β > 1. Taking the limit of the last expression as k we obtain 11 1β 3β 1x1 + β 5 β + 3β β 3 x1 + β + 1 + ln1/x if β 1, + 1 + ln1/x if β > 1. The maximum of the last expression is obtained at x = 11+1β+3β 11+β when β 1/ and at x = when β > 1/ in which case the expression amounts to 5+β 3β +β 3 1+β 11 + β ln 11 + 1β + 3β 1 + β ln 5 + β 3β + β 3 18 if β 1, if β > 1.

This proves that 11 + β ψ ln 11 + 1β + 3β ψ ln 1 + β 5 + β 3β + β 3 n 1 o n 1 if β 1, n 1 o n 1 if β > 1. This completes the proof of the lower bound in Theorem 1. which, together with the upper bound proved in Section, yields the entire proof of Theorem 1.. Acknowledgment We thank the two anonymous referees for useful comments improving the presentation of the paper. References [1] N. Alon and A. Shapira. Testing subgraphs in directed graphs. In Proceedings of the thirty-fifth annual ACM symposium on Theory of computing, pages 700 70. ACM, 003. [] M. Chudnovsky. The Erdös-Hajnal conjecture - a survey. Journal of Graph Theory, 75:178 10, 014. [3] A. Czygrinow, S. Poljak, and V. Rödl. Constructive quasi-ramsey numbers and tournament ranking. SIAM Journal on Discrete Mathematics, 11:48 3, 1. [4] R.B. Gardner. Optimal packings and coverings of the complete directed graph with 3-circuits and with transitive triples. Congressus Numerantium, pages 11 170, 17. [5] P.E. Haxell and V. Rödl. Integer and fractional packings in dense graphs. Combinatorica, 11:13 38, 001. [] M. Kabiya and R. Yuster. Packing transitive triples in a tournament. Annals of Combinatorics, 13:1 30, 008. [7] P. Keevash and B. Sudakov. Triangle packings and 1-factors in oriented graphs. Journal of Combinatorial Theory, Series B, 4:70 77, 00. [8] T.P. Kirkman. On a problem in combinatorics. Cambridge and Dublin Mathematical Journal, :11 04, 1847. [] D. Kühn and D. Osthus. Hamilton decompositions of regular expanders: A proof of kellys conjecture for large tournaments. Advances in Mathematics, 37: 14, 013. 1

[10] B.D. McKay. The asymptotic numbers of regular tournaments, eulerian digraphs and eulerian oriented graphs. Combinatorica, 104:37 377, 10. [11] Z. Nutov and R. Yuster. Packing directed cycles efficiently. Discrete Applied Mathematics, 155:8 1, 007. [1] K.T. Phelps and C.C. Lindner. On the number of mendelsohn and transitive triple systems. European Journal of Combinatorics, 53:3 4, 184. [13] P. Rowlinson. On 4-cycles and 5-cycles in regular tournaments. Bulletin of the London Mathematical Society, 18:135 13, 18. [14] E. Szemerédi. Regular partitions of graphs. In Problèmes combinatoires et théorie des graphes Colloq. Internat. CNRS, Univ. Orsay, Orsay, 17, volume 0 of Colloq. Internat. CNRS, pages 3 401. CNRS, Paris, 178. [15] R. Yuster. Packing triangles in regular tournaments. Journal of Graph Theory, 741:58, 013. 0